Loading...
Loading...
Dataset·Catalogue
A growing class of models never writes a sentence. Hand one a state and a typed question and it returns a choice, a score or a probability, ready for code to act on. This catalogue tracks them, from the new general decision models to the rerankers and guardrails already in production, and records what each source actually claims about whether those probabilities can be trusted.
Decision Models carries no numeric fields -- showing its fact columns instead, and no median footer (a median of text is not a median).
| # | ||||||||
|---|---|---|---|---|---|---|---|---|
| 1 | BGE Reranker v2-M3 [Open weights] | ? | BAAI (Beijing Academy of Artificial Intelligence) | Open weights | none claimed | apache-2.0 | not published | not published |
| 2 | Clef [API and open weights] | 2026-10-01 | Cloudflare | API and open weights | claimed | Apache-2.0 | $0.24 per M input tokens on Workers AI; Clef models do not charge for output tokens. | 65,536 tokens |
| 3 | Clef-flash [API and open weights] | 2026-10-01 | Cloudflare | API and open weights | claimed | Apache-2.0 | $0.038 per M input tokens on Workers AI; Clef models do not charge for output tokens. | 24,576 tokens |
| 4 | Cohere Rerank (v4.0 pro/fast, v3.5, v3.0) [Hosted API] | ? | Cohere | Hosted API | none claimed | not applicable | not published | not published |
| 5 | d1-3B (with d1-omni-600M) [Open weights] | 2026-10-07 | Liquid AI | Open weights | claimed | LFM Open License (lfm1.0) | not published | 32,768 tokens |
| 6 | Gaggle (pointwise decoder-only reranker) [Open weights] | 2026-10-09 | Castorini (University of Waterloo) | Open weights | none claimed | MIT | not published | backbone supports a 2,048-token context; trained with 128 query tokens and 512 passage tokens, stated as not a hard limit at inference |
| 7 | HHEM-2.1-Open (Hughes Hallucination Evaluation Model) [Open weights] | ? | Vectara | Open weights | none claimed | apache-2.0 | not published | unlimited context length (vendor-stated; HHEM-1.0 was capped at 512 tokens) |
| 8 | Horizon Labs Multilingual Reranker (base, 308M) [Open weights] | 2026-10-01 | Horizon Labs | Open weights | none claimed | apache-2.0 | not published | trained with 384 tokens per pair; longer passages are truncated |
| 9 | Jev 1.13 [Hosted API] | ? | TypeSafe | Hosted API | claimed | not applicable | $42 per billion input tokens ($0.042 per million). Charged per input token; output tokens are free. | 64k tokens per request; 32k tokens for state plus the longest question. Both rise to 262k with the TypeSafe-Danger-Zone-Long-Context header, which TypeSafe says is slower and less accurate. |
| 10 | Jev 1.14 [Hosted API] | 2026-10-14 | TypeSafe | Hosted API | claimed | not applicable | $42 per billion input tokens ($0.042 per million). Output tokens are free. | 64k tokens per request; 32k tokens per question. Both rise to 262k with the TypeSafe-Danger-Zone-Long-Context header, which TypeSafe says is slower and less accurate. |
| 11 | Jina Classifier API [Hosted API] | ? | Jina AI | Hosted API | none claimed | not applicable | API pricing is based on the token usage. | not published |
| 12 | jina-reranker-v3.5 [API and open weights] | ? | Jina AI | API and open weights | none claimed | CC-BY-NC-4.0 (commercial use requires contacting Jina AI) | not published | 131K tokens |
| 13 | Llama Prompt Guard 2 (86M) [Open weights] | ? | Meta | Open weights | none claimed | Llama 4 Community License | not published | not published |
| 14 | Llama-3.1-Nemotron-70B-Reward [API and open weights] | ? | NVIDIA | API and open weights | none claimed | NVIDIA Open Model License plus Llama 3.1 Community License Agreement | NVIDIA offers free hosted inference at build.nvidia.com with an OpenAI-compatible API interface; signing up includes 100k free API calls to this model. | up to 4,096 tokens |
| 15 | Microsoft-Decision-1 [Hosted API] | ? | Microsoft | Hosted API | claimed | not applicable | not published | 32,768 tokens |
| 16 | mxbai-rerank-large-v2 [Open weights] | ? | Mixedbread | Open weights | none claimed | apache-2.0 | not published | not published |
| 17 | omni-moderation-latest [Hosted API] | ? | OpenAI | Hosted API | none claimed | not applicable | the moderation endpoint is free to use | not published |
| 18 | pplx-decider v1.2 (with flash-v1.2) [Open weights] | ? | Perplexity | Open weights | claimed | Apache-2.0 | not published | max_length defaults to 262,144 tokens per sequence (the backbone position limit); training used sequences up to about 32k tokens |
| 19 | Qwen3Guard-Stream-4B [Open weights] | ? | Alibaba Qwen | Open weights | none claimed | apache-2.0 | not published | not published |
| 20 | ShieldGemma 2B [Open weights] | 2024-07-21 | Open weights | none claimed | Gemma license (Google usage terms) | not published | not published | |
| 21 | Strands decider (gemma4 v1-2610 family and 2B Qwen line) [Open weights] | 2026-10-09 | Strands Agents (strands-labs) | Open weights | claimed | Apache-2.0 | not published | not published |
| 22 | Trustworthy Language Model (TLM) score-only mode [Hosted API] | ? | Cleanlab | Hosted API | claimed | not applicable | not published | not published |
| 23 | Vertex AI Ranking API (semantic ranker) [Hosted API] | 2025-04-09 | Google Cloud (Vertex AI Agent Builder) | Hosted API | none claimed | not applicable | not published | not published |
| 24 | Voyage Rerank (rerank-3 family; rerank-2.5 now legacy) [Hosted API] | ? | Voyage AI (acquired by MongoDB, February 2025) | Hosted API | none claimed | not applicable | not published | not published |